Re: [PATCH net 1/2] bpf, sockmap: Avoid sk_prot reset on sockmap unlink with ULP set

[Cong Wang <xiyou.wangcong@gmail.com>]

On Fri, Mar 14, 2025 at 04:20:03PM +0800, Dong Chenchen wrote:
> WARNING: CPU: 0 PID: 6558 at net/core/sock_map.c:1703 sock_map_close+0x3c4/0x480
> Modules linked in:
> CPU: 0 UID: 0 PID: 6558 Comm: syz-executor.14 Not tainted 6.14.0-rc5+ #238
> RIP: 0010:sock_map_close+0x3c4/0x480
> Call Trace:
>  <TASK>
>  inet_release+0x144/0x280
>  __sock_release+0xb8/0x270
>  sock_close+0x1e/0x30
>  __fput+0x3c6/0xb30
>  __fput_sync+0x7b/0x90
>  __x64_sys_close+0x90/0x120
>  do_syscall_64+0x5d/0x170
>  entry_SYSCALL_64_after_hwframe+0x76/0x7e
> 
> The root cause is:
> bpf_prog_attach(BPF_SK_SKB_STREAM_VERDICT)
> tcp_set_ulp //set ulp after sockmap add
> 	icsk->icsk_ulp_ops = ulp_ops;
> sock_hash_update_common
>   sock_map_unref
>     sock_map_del_link
>       psock->psock_update_sk_prot(sk, psock, false);
> 	sk->sk_prot->close = sock_map_close
> sk_psock_drop
>   sk_psock_restore_proto
>     tcp_bpf_update_proto
>        tls_update //not redo sk_prot to tcp prot
> inet_release
>   sk->sk_prot->close
>     sock_map_close
>       WARN(sk->sk_prot->close == sock_map_close)

This makes sense now. Please see my comment below.

> 
> commit e34a07c0ae39 ("sock: redo the psock vs ULP protection check")
> has moved ulp check from tcp_bpf_update_proto() to psock init.
> If sk sets ulp after being added to sockmap, it will reset sk_prot to
> BPF_BASE when removed from sockmap. After the psock is dropped, it will
> not reset sk_prot back to the tcp prot, only tls context update is
> performed. This can trigger a warning in sock_map_close() due to
> recursion of sk->sk_prot->close.
> 
> To fix this issue, skip the sk_prot operations redo when deleting link
> from sockmap if ULP is set.
> 
> Fixes: e34a07c0ae39 ("sock: redo the psock vs ULP protection check")
> Fixes: c0d95d3380ee ("bpf, sockmap: Re-evaluate proto ops when psock is removed from sockmap")
> Signed-off-by: Dong Chenchen <dongchenchen2@huawei.com>
> ---
>  net/core/sock_map.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/net/core/sock_map.c b/net/core/sock_map.c
> index 82a14f131d00..a3ed1f2cf8a2 100644
> --- a/net/core/sock_map.c
> +++ b/net/core/sock_map.c
> @@ -170,7 +170,7 @@ static void sock_map_del_link(struct sock *sk,
>  		if (verdict_stop)
>  			sk_psock_stop_verdict(sk, psock);
>  
> -		if (psock->psock_update_sk_prot)
> +		if (!(sk_is_inet(sk) && inet_csk_has_ulp(sk)) && psock->psock_update_sk_prot)
>  			psock->psock_update_sk_prot(sk, psock, false);

Can we put this TCP-specific logic into tcp_bpf_update_proto() instead?

Something like this...

diff --git a/net/ipv4/tcp_bpf.c b/net/ipv4/tcp_bpf.c
index ba581785adb4..0bb363447fc7 100644
--- a/net/ipv4/tcp_bpf.c
+++ b/net/ipv4/tcp_bpf.c
@@ -708,6 +708,8 @@ int tcp_bpf_update_proto(struct sock *sk, struct sk_psock *psock, bool restore)
                return 0;
        }

+       if (inet_csk_has_ulp(sk))
+               return 0;
        if (sk->sk_family == AF_INET6) {
                if (tcp_bpf_assert_proto_ops(psock->sk_proto))
                        return -EINVAL;